Diagnostic test accuracy of AI-assisted mammography for breast imaging: a narrative review

Data diversity and quality

The success of AI algorithms in mammography is profoundly intertwined with the quality and diversity of the data used for training. An AI model’s ability to accurately identify abnormalities and interpret mammographic images is contingent upon exposure to a comprehensive range of breast conditions, encompassing diverse anatomical variations, breast compositions, and demographic factors (Jairam & Ha, 2022). However, assembling a dataset that genuinely mirrors the complex spectrum of breast characteristics presents a formidable challenge. Historically, medical imaging datasets have often skewed towards certain demographic groups, potentially introducing bias and limiting the generalizability of AI algorithms (Sheth & Giger, 2020). Addressing this challenge requires the collection of data that comprehensively represents the mosaic of breast types encountered in clinical practice, encompassing varying breast densities, ages, ethnicities, and pathological conditions.

Achieving data diversity extends beyond sheer numbers; it requires meticulous curation and annotation to ensure that the dataset faithfully reflects the realities of mammographic interpretations (Mehrotra et al., 2020; Halling-Brown et al., 2020). Efforts to overcome the challenge of data diversity involve collaborating with healthcare institutions to access a broad patient base and partnering with radiologists to annotate images accurately. Leveraging data augmentation techniques can help expand the dataset’s diversity by simulating various imaging scenarios and breast compositions. Moreover, data quality is paramount for robust AI model training. The dataset must be meticulously quality-controlled to eliminate artifacts, ensure proper image alignment, and standardize image acquisition protocols. Inconsistencies in data quality can compromise the AI algorithm’s ability to learn relevant patterns and, subsequently, its diagnostic accuracy.

Solving the data diversity and quality challenge requires a concerted effort from the medical imaging community, healthcare providers, and AI researchers (Halling-Brown et al., 2020). Collaboration in assembling diverse and representative datasets, addressing bias, and adopting standardized imaging protocols is pivotal for cultivating AI models that can be effectively integrated into clinical practice. Robust data diversity and quality not only enhance AI’s diagnostic accuracy but also lay the foundation for equitable healthcare outcomes across diverse patient populations.

Data diversity and skewed datasets are critical concerns in artificial intelligence (AI), especially regarding model performance and generalizability.

Skewed datasets refer to datasets where one class or group is disproportionately represented compared to others. This imbalance can lead AI models to overfit on the majority class while underperforming on minority classes. Balashankar et al. (2019) discuss a Pareto-efficient fairness approach to handle skewed subgroup data, showing that improving performance across all subgroups is challenging in skewed datasets. Seyyed-Kalantari et al. (2021) explored underdiagnosis bias in AI systems applied to medical images, particularly in underserved patient populations. They found that skewed datasets often lead to poor performance in minority groups, which can have severe implications in healthcare applications. Törnquist & Caulk (2024) explore how curating diverse, synthetic data can enhance AI performance, particularly in skewed datasets. Their work shows that enriching datasets with strategic diversification improves model generalizability.

To improve the performance of AI in mammography and ensure equitable healthcare outcomes, it is crucial to include geographically diverse datasets in the training process. Incorporating mammograms from various regions can help AI models generalize better across different populations and healthcare systems. A more geographically diverse dataset accounts for variations in imaging devices, protocols, and patient demographics, ensuring that the models are robust and adaptable. Recent efforts to compile large-scale, diverse datasets, such as the OPTIMAM Mammography Image Database (OMI-DB) (OMI-DB, 2020), which includes mammograms from different countries and populations, have shown promise in addressing these challenges. Expanding the geographical diversity of AI training datasets can contribute to the development of more accurate and reliable models that can be effectively deployed in various clinical settings worldwide .

Generalization

The robustness of AI algorithms in medical imaging, including AI-assisted mammography, hinges on their capacity to generalize findings from controlled datasets to real-world clinical scenarios. While AI models might excel in specific datasets, ensuring their consistent performance across diverse clinical conditions is a critical challenge (Barnett et al., 2021).

Real-world mammography encounters a myriad of variables that can impact image quality and interpretation. Imaging devices vary in terms of resolution, noise levels, and imaging protocols (Lauritzen et al., 2022). Patient characteristics, such as breast density and anatomical variations, introduce additional complexity. Moreover, the clinical context, encompassing various pathologies and disease stages, can significantly affect mammographic appearance. For AI-assisted mammography to be clinically effective, it must transcend the limitations of dataset-specific learning and demonstrate generalizability (Lauritzen et al., 2022). Achieving robust generalization requires exposing AI algorithms to a diverse range of clinical scenarios during training. Incorporating images from various institutions, acquired using different imaging devices and protocols, is pivotal. Augmenting the training dataset with synthetic variations can further enhance the AI model’s ability to adapt to real-world challenges.

The challenge of generalization also necessitates continuous monitoring and validation post-deployment. AI algorithms should be rigorously tested across multiple clinical settings to ensure consistent performance. Feedback loops, where AI-generated interpretations are compared against ground truth assessments by radiologists, play a vital role in fine-tuning models and addressing any performance disparities that arise in specific contexts. Furthermore, addressing bias in AI algorithms is intertwined with achieving robust generalization. Bias that emerges from imbalanced datasets or algorithmic tendencies can result in suboptimal performance in certain patient populations or clinical conditions. Regularly auditing algorithms for bias and recalibrating them to mitigate biases are essential steps in enhancing generalization and ensuring equitable healthcare outcomes. In the pursuit of effective generalization, collaboration between AI researchers, radiologists, and healthcare institutions is paramount. A diverse training dataset, continuous validation, and ongoing refinement of AI models contribute to their ability to navigate the intricacies of real-world mammography and provide reliable support to clinicians.

Interpretability

The integration of AI into medical imaging, such as AI-assisted mammography, brings to the forefront the challenge of interpretability - the ability to understand and explain the decisions made by AI algorithms (Killock, 2020). Interpretability is of paramount importance in clinical settings where radiologists, clinicians, and patients seek transparency in diagnostic processes. The intricate nature of AI algorithms, particularly those based on deep learning, often results in a “black-box” phenomenon, where the rationale behind algorithmic decisions is not readily apparent (Killock, 2020). This lack of transparency raises concerns about trust, accountability, and the ability to validate AI-generated diagnoses. Radiologists and clinicians need to have confidence not only in the accuracy of AI predictions but also in the insights that underpin those predictions.

Addressing the challenge of interpretability involves developing methods that shed light on the inner workings of AI models. Researchers are exploring techniques to visualize the features and patterns that AI algorithms focus on when making decisions (Thrall, Fessell & Pandharipande, 2021). These visualizations can help radiologists understand the reasoning behind AI-generated findings and enable them to integrate AI interpretations seamlessly into their clinical workflows.

Many current AI models, particularly deep learning algorithms, are often considered “black boxes” due to their complex and opaque nature. This lack of transparency is a significant barrier to clinical adoption. Clinicians require interpretable models to understand the decision-making process behind AI predictions, especially when the stakes involve patient health and treatment outcomes (Jia, Ren & Cai, 2020).

Another approach to enhance interpretability involves developing “explainable AI” methods. These methods aim to generate explanations that are comprehensible to humans, providing insight into the factors that influenced an AI decision. Techniques like feature importance scores, attention maps, and saliency maps offer avenues for presenting AI-generated results in a manner that resonates with the expertise of radiologists. Furthermore, efforts to address interpretability intersect with the broader theme of trust-building in AI-assisted mammography. When radiologists and clinicians can clearly comprehend how AI arrives at its conclusions, they are more likely to trust and embrace AI-generated insights as valuable complements to their own expertise (Leibig et al., 2022; Rodriguez-Ruiz et al., 2019a). However, striking a balance between interpretability and performance remains a challenge. Simplifying AI models for better interpretability might lead to a trade-off in predictive accuracy. Achieving both high performance and explainability requires innovative research into model architectures and interpretability techniques.

In the pursuit of enhancing interpretability, collaboration between AI researchers, radiologists, and ethicists is essential. Transparent and interpretable AI algorithms hold the potential to not only improve diagnostic accuracy but also foster a cooperative synergy between AI systems and human experts.

Regulatory and ethical concerns

The integration of AI into medical imaging, particularly AI-assisted mammography, introduces a host of regulatory and ethical considerations that are essential to navigate in order to ensure patient safety, data privacy, and equitable healthcare delivery (Aggarwal et al., 2021; Kleppe et al., 2021). The deployment of AI in clinical settings involves a careful balance between innovation and adherence to established regulatory frameworks.

Validation and benchmarking

The integration of AI into medical imaging, including AI-assisted mammography, necessitates a rigorous process of validation and benchmarking to ensure the reliability and reproducibility of AI-generated results (Yala et al., 2022). The validation framework serves as a critical bridge between algorithm development and clinical implementation, guiding the assessment of AI performance and enhancing its clinical utility.

Inclusion criteria

A rigorous set of inclusion and exclusion criteria was established to guide the selection of studies for this review. The inclusion criteria were carefully designed to ensure the alignment of selected studies with the review’s central focus on the diagnostic test accuracy of AI-assisted mammography. The inclusion criteria were as follows:

Relevance to Diagnostic Accuracy: Studies that explicitly examined the diagnostic test accuracy of AI algorithms applied to mammographic imaging were considered eligible for inclusion. Human Subjects: To capture the real-world clinical application of AI-assisted mammography, only studies involving human subjects were included. Diagnostic Accuracy Metrics: Studies that reported quantitative diagnostic accuracy metrics, including sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC), or other relevant metrics, were included. Publication Language: To ensure accessibility and feasibility for the review authors and readers, only studies published in the English language were included.

We have established threshold for diagnostic accuracy metrics using AUC-ROC, as a primary criterion for study inclusion. We have included only those studies (literatures) that has AUC-ROC >80%.

Exclusion criteria

Exclusion criteria were designed to refine the selection process and maintain the study’s focus on diagnostic accuracy in AI-assisted mammography. The exclusion criteria were as follows:

Non-Diagnostic Accuracy Studies: Studies that did not primarily investigate diagnostic accuracy, such as those exploring technical aspects of AI algorithms or involving non-human subjects, were excluded. Non-English Publications: Studies published in languages other than English were excluded due to potential language barriers and the available resources of the review.

Foundations of AI in mammography: unraveling the algorithmic Marvel

AI-assisted mammography is an epitome of technological ingenuity, where machine learning algorithms, particularly convolutional neural networks (CNNs), decode the complexity of mammograms (Wu et al., 2019; Gastounioti et al., 2022). These algorithms have the capacity to recognize intricate patterns and minute features that often elude human perception. By being trained on extensive datasets of annotated mammograms, AI algorithms learn to discern subtle hallmarks of potential malignancies, providing a foundation to augment diagnostic accuracy and mitigate the risks of misdiagnosis.

Diverse AI approaches: navigating a multiverse of techniques

Within the spectrum of AI-assisted mammography, diverse strategies converge to address the multifaceted challenges of breast cancer detection. Ranging from computer-aided detection (CAD), which acts as a vigilant spotlight on regions of interest, to the diagnostic prowess of computer-aided diagnosis (CADx), providing quantitative insights, the possibilities are boundless (Katzen & Dodelzon, 2018; Vyborny & Giger, 1994). This adaptability extends across various mammographic modalities, encompassing the established full-field digital mammography (FFDM), the emerging digital breast tomosynthesis (DBT), and even experimental modalities such as contrast-enhanced mammography.

Training AI models: nurturing intelligence in silicon

The essence of AI’s metamorphosis lies in its capacity to learn from data at a scale and complexity unattainable by human endeavor. Training AI models for mammography entails immersing them in meticulously annotated datasets (Mayo et al., 2019). Each image serves as a repository of diagnostic outcomes, enabling the algorithms to decipher the intricate tapestry of patterns and pathologies. Through iterative refinement, these models internalize mammographic nuances, unraveling even the most nuanced signs of malignancy (Rodríguez-Ruiz et al., 2019). The crescendo of this process culminates in rigorous validation, refining their diagnostic acumen.

Xia et al. (2023) discuss the enhanced moth-flame optimizer with quasi-reflection and refraction learning, demonstrating its application in image segmentation and medical diagnosis. Jung et al. (2024a) conducts a systematic review and meta-analysis on the use of AI in fracture detection across various image modalities and data types, published in PLOS Digital Health. Fan et al. (2023) explore the clinical characteristics, diagnosis, and management of Sweet syndrome induced by azathioprine, providing crucial insights into autoimmune responses and drug interactions. Zhu (2024) focuses on a computational intelligence-based classification system for diagnosing memory impairment in psychoactive substance users, offering novel approaches to mental health diagnostics.

 He et al. (2020) presents a new method for recognizing circulating tumor cells (CTC) using machine learning, underscoring significant strides in cancer detection. Hu et al. (2024) propose a trustworthy multi-phase liver tumor segmentation method via evidence-based uncertainty, enhancing accuracy in liver cancer diagnostics. In a study by Hu et al. (2023), the accuracy of Gallium-68 Pentixafor PET-CT for subtyping primary aldosteronism is evaluated, published in JAMA Network Open. Yang et al. (2024) introduces a dual-domain diffusion model for sparse-view CT reconstruction, pushing the boundaries of imaging technology. Lastly, Yin et al. (2024) detail the use of a convolution-transformer for image feature extraction, highlighting innovative approaches in engineering and AI integration in medical imaging.

Diagnostic enhancement potential: envisioning uncharted diagnostic horizons

The promise of AI in mammography resonates with its potential to elevate diagnostic accuracy. Collaborating with radiologists’ clinical expertise, AI algorithms forge a symbiotic alliance. This partnership amplifies diagnostic precision, mitigates the risks of diagnostic errors, and opens pathways to early intervention (Lång et al., 2021b; Yoon & Kim, 2021). AI algorithms serve as vigilant co-pilots, offering an additional layer of scrutiny that illuminates subtle regions of concern, bridging the gap between human visual acuity and computational prowess. However, this integration necessitates meticulous validation, seamless clinical incorporation, and ongoing collaboration to ensure the fusion of human and AI insights.

Challenges and considerations: navigating the labyrinthine terrain

As AI ushers in a new era, challenges are woven into its fabric. The intricate diversity of breast tissue composition, the nuances of image quality across diverse modalities, and the aspiration for algorithms to generalize across diverse populations present formidable hurdles (Bahl, 2019; Lång et al., 2021a). Ethical considerations loom large, encompassing data privacy, algorithmic bias, and the imperative to address potential underrepresentation. Harmonizing AI into clinical workflows necessitates a seismic shift, incorporating interpretability, accountability, and mechanisms for human oversight.

Prospects and conclusion: charting uncharted territories

The symbiotic alliance of AI and mammography resounds beyond technology, reverberating through healthcare’s core. This section, “AI-Assisted Mammography: An Overview”, propels us into an expedition that delves deep into AI’s role in refining diagnostic accuracy. As we navigate through a comprehensive review of studies, we uncover insights illuminating AI’s potential to reshape breast imaging’s landscape.

The realm of AI-assisted mammography unveils a constellation of studies, each bearing the potential to reshape breast cancer diagnosis. This literature review journey traverses the breadth of 13 distinct studies, extracting their quintessence to weave a comprehensive narrative that highlights the multifaceted dimensions of diagnostic test accuracy in AI-assisted mammography as shown in Fig. 1.

Yala et al. (2019) illuminate the path with their deployment of ResNet-18, a neural network renowned for its convolutional architecture. Their exploration, nested within a monumental dataset of 212,276 cases in the United States, yielded a luminous sensitivity of 90.1% and a specificity of 94.2%, unveiling the profound capacity of AI to discern subtleties imperceptible to the human eye (Yala et al., 2019).

McKinney et al. (2020) embark on a computational odyssey, employing an ensemble of ResNet models, MobileNetV2, and RetinaNet. Their intricate dance of algorithms transcended boundaries, encompassing the United Kingdom and the United States. Amidst a canvas of 26,142 cases, their results unfolded with a sensitivity of 65.42% and a specificity of 94.12%, underscoring AI’s adaptability to distinct healthcare landscapes (McKinney et al., 2020).

Dembrower et al. (2020) channel the power of Lunit version 5.5.0, carving a unique trajectory through 170,230 cases in Sweden. Their exploration ventured beyond raw accuracy, delving into the realm of missed cancers across varying sensitivities. This nuanced examination reframes AI’s role not merely as an analytical tool, but as a means to scrutinize diagnostic reliability (Dembrower et al., 2020).

Kyono et al., in their dual explorations of 2018 (Kyono, Gilbert & Van der Schaar, 2018) and 2020 (Kyono, Gilbert & Van der Schaar, 2020), intertwine diagnostic metrics with the fabric of machine learning. The UK-based study of 2018 harnessed InceptionResNetV2, revealing metrics such as Cohen’s k of 0.716 and an F1 statistical test score of 0.757. The subsequent chapter, cast in 2020, is marked by an exploration of NPV beyond 99.0%. Together, they reflect AI’s dual facets as an analytical instrument and a safeguard of diagnostic precision (Kyono, Gilbert & Van der Schaar, 2020).

Rodriguez-Ruiz et al. (2019b) orchestrate a symphony spanning multiple countries, courtesy of Transpara (version 1.6.0). Their contribution, punctuated by a dataset of 189,000 cases, narrates a tale of sensitivity at 83.0% and specificity at 77.0%. This cross-continental panorama underscores AI’s potential to harmonize with diverse healthcare ecosystems.

Lotter et al. (2021) invoked annotation efficiency through ResNet-50 and RetinaNet. Against the backdrop of 97,769 cases in the United States, their journey resonates with sensitivities of 96.2% and specificities of 90.9%, underlining AI’s promise in enriching diagnostic accuracy.

Schaffter et al. (2020) conduct a symposium of networks, harmonizing CEM and VGG ensembles. Their exploration unfurls across the United States and Sweden, celebrating sensitivities of 85.9% and specificities of 88.0%. This narrative within networks captures the essence of balanced diagnostic acumen.

Salim et al. (2020) grace Sweden’s healthcare canvas with ResNet-34 and MobileNet. Amidst the voluminous expanse of 752,000 cases, they unveil a sensitivity of 81.9% intertwined with a specificity of 96.6%, embodying AI’s potential to bolster diagnostic fidelity.

Kim et al. (2020), in their Korean journey, wield ResNet-34 to unravel sensitivities of 88.8% and specificities of 81.9%. Their exploration fuses machine learning’s analytical gaze with South Korea’s diagnostic narrative.

Taylor et al. (2005) paints a tale of human-AI symbiosis, illuminating the United Kingdom with computer-aided detection prompts. This narrative oscillates with a sensitivity of 76.0% and specificity of 83.0%, reflecting the fine dance of human expertise and AI’s analytical precision.

Fenton et al. (2011) undertake an opus across 1.6 million images in the United States, revealing sensitivities of 85.3% and specificities of 91.4%. Their contribution punctuates the landscape with benchmarks for AI’s diagnostic aspirations.

Becker et al. (2017) add a Swiss touch with ViDi Suite Version 2.0, etching sensitivities of 73.7% and specificities of 72.0%. This Swiss narrative enriches the mosaic of global AI’s diagnostic endeavors.

Collectively, these studies converge as luminous stars in the cosmos of AI-assisted mammography. They beckon forth a landscape imbued with AI’s potential to elevate diagnostic accuracy, reshape healthcare narratives, and unfurl a new dawn in breast cancer diagnosis as reported in Table 1.

Approaches used: unleashing architectural marvels

Across the spectrum of studies, algorithmic techniques emerge as a critical theme, showcasing the power of deep learning architectures. ResNet-18, a convolutional neural network, stands tall in the work of Yala et al. (2019), yielding a remarkable sensitivity of 90.1% and specificity of 94.2% in the United States. InceptionResNetV2, embraced by Kyono, Gilbert & Van der Schaar (2018) and Kyono, Gilbert & Van der Schaar (2020), unravels metrics such as Cohen’s k of 0.716 and an F1 statistical test score of 0.757, highlighting the potential for nuanced diagnostic assessment. These architectural marvels lay the groundwork for AI’s role in amplifying mammography’s diagnostic accuracy as reported in Table 2.

Geographic contexts: bridging healthcare landscapes

The thematic tapestry extends to geographic contexts, where AI transcends boundaries to resonate in distinct healthcare landscapes. McKinney et al. (2020) exemplify this theme, deploying ensemble ResNet models, MobileNetV2, and RetinaNet to achieve a sensitivity of 65.42% and specificity of 94.12% across the United Kingdom and the United States. Schaffter et al. (2020) harmonize CEM and VGG networks, juxtaposing sensitivities of 85.9% and specificities of 88.0% between the United States and Sweden. Such cross-continental explorations underscore AI’s potential to bridge diagnostic challenges inherent to diverse healthcare ecosystems.

Diagnostic metrics: the language of accuracy

Diagnostic metrics resonate as a central theme, capturing the essence of AI’s diagnostic augmentation. Lotter et al. (2021) utilize ResNet-50 and RetinaNet to unveil sensitivities of 96.2% and specificities of 90.9% in the United States. Salim et al. (2020) infuse Sweden’s healthcare fabric with ResNet-34 and MobileNet, attaining a sensitivity of 81.9% and specificity of 96.6%. These diagnostic metrics not only quantify AI’s performance but also articulate its potential in recalibrating breast cancer diagnosis.

Artificial intelligence based models in mammography

Deep learning, a subset of AI, has gained significant traction in mammography due to its ability to analyze large datasets and identify patterns that may go unnoticed by human radiologists. Convolutional neural networks (CNNs), in particular, are widely used for their capacity to detect breast abnormalities such as masses, calcifications, and architectural distortions. In 2022, Gastounioti et al. (2022) trained to phenotype breast cancer risk using mammographic features. They observed that AI could significantly improve the risk assessment by capturing subtle imaging biomarkers, outperforming traditional risk models in predicting breast cancer development. Yin et al. (2024) focused on using a convolution-transformer for image feature extraction, which improved the accuracy of mammographic image interpretation. The integration of AI systems into radiology workflows helped radiologists identify breast cancer in earlier stages, reducing false negatives and potentially improving patient outcomes.

Machine learning techniques are also being employed to enhance diagnostic accuracy in mammography. These models learn from large datasets of mammograms and can assist radiologists by suggesting areas of concern, which helps in reducing human error. Zhu (2024) explored how AI-based classification systems could improve the diagnosis of breast cancer using mammography images. The machine learning models were trained using a variety of breast images and patient data to accurately classify different types of breast abnormalities. The system achieved high sensitivity and specificity rates, indicating that AI can help improve early detection and reduce the burden on radiologists. Moreover, Jung et al. (2024a) conducted a systematic review and meta-analysis of AI systems in medical imaging, including mammography, showing that AI tools significantly improved the diagnostic performance of radiologists when used as decision support tools. This research highlights the growing role of AI in clinical settings where it enhances diagnostic capabilities and reduces the time required to interpret images.

AI models have been particularly effective in addressing the issue of false positives and false negatives in mammography, which can lead to unnecessary biopsies or missed cancer diagnoses. Salim et al. (2020) conducted an external evaluation of commercial AI algorithms for mammogram assessment and found that AI could independently assess mammograms and reduce the false-positive rate without compromising cancer detection rates. Similarly, He, Wang & Zhang (2023) introduced a novel method for tumor detection in mammography using machine learning. Their method improved upon previous algorithms by reducing false-negative rates, particularly in patients with dense breast tissue. Dense tissue often obscures tumors, making early detection more challenging, but AI systems were shown to provide greater accuracy in these difficult cases. Hu et al. (2024) highlighted the need for trustworthy AI systems in mammography, advocating for evidence-based approaches to AI model development that prioritize transparency and interpretability. They emphasized the importance of designing AI systems that radiologists and healthcare providers can trust, especially when patients’ lives depend on accurate diagnoses. In addition, Zhu (2024) pointed out the risk of bias in AI models that are trained on non-diverse datasets. Since most training data come from developed regions, the algorithms may not generalize well to populations in developing countries, potentially leading to disparities in diagnostic outcomes.

Fan et al. (2023) suggested that integrating AI with other healthcare technologies, such as electronic health records (EHRs) and genetic testing, could lead to more personalized and precise breast cancer screening protocols. AI-driven insights from mammograms could be combined with genetic risk factors to provide a holistic view of a patient’s breast cancer risk, thereby enabling personalized treatment plans.

Comparison between approaches used for AI-assisted mammography of breast imaging

Table 3 has reported the comparison between different approaches used for AI-assisted mammography for breast imaging.

Data diversity and quality

To enhance the accuracy and fairness of AI models in mammography, it is essential to foster collaborative data collection by encouraging partnerships between healthcare institutions globally. This approach will help build comprehensive datasets that encompass diverse demographics, breast compositions, and pathological variations, ultimately reducing bias in AI systems (Dlamini et al., 2020). Additionally, implementing data augmentation techniques, such as simulating different imaging scenarios, will artificially expand dataset diversity, ensuring that AI models can generalize across a broader range of clinical conditions. Furthermore, collaboration across institutions to standardize image acquisition protocols is crucial for ensuring consistency in data collection, which will reduce noise in AI training data and improve model performance.

Generalization of AI models

To ensure AI models in mammography perform effectively in real-world scenarios, it is crucial to incorporate multi-institutional data by using datasets from diverse geographic locations, imaging devices, and clinical settings (Ginsburg et al., 2020). This approach enhances the robustness of algorithms beyond controlled environments. Additionally, continuous monitoring post-deployment is essential, with feedback loops enabling radiologists to provide real-time input on AI performance in clinical practice. This ongoing feedback allows AI systems to learn and adapt to new conditions, ensuring they remain accurate and relevant over time.

Interpretability

To facilitate the integration of AI into clinical workflows, it is essential to develop explainable AI methods, such as attention maps, saliency maps, and feature importance scores, which help radiologists understand why AI algorithms make specific predictions. These tools will enhance transparency and increase confidence in AI-driven decisions (Batchu et al., 2021). Additionally, simplifying model architectures can strike a balance between complexity and interpretability, making AI systems more transparent to radiologists and avoiding the “black-box” problem, where the inner workings of the model are not easily understood by end users.